Metraipolis › archive › dc_94f86c8b6441

Deterministic Fact Validation for Retrieval-Augmented Generation of Clinical Data Reviewer's Guides: A Proposed Architecture (2026) · DOI 10.21203/rs.3.rs-10884117/v1

Source
Metraipolis archive
Original
https://doi.org/10.21203/rs.3.rs-10884117/v1
License (as archived)
CC-BY-4.0
Archived
2026-10-07T09:12:38.558Z
SHA-256 of the archived text
4d67b3ddb86ce9eef6dc8510bb8b8ed2216e00133268bb84e66355012091a161
Stable URL of this page
https://metraipolis.com/documento/dc_94f86c8b6441

How to cite

Deterministic Fact Validation for Retrieval-Augmented Generation of Clinical Data Reviewer's Guides: A Proposed Architecture (2026) · DOI 10.21203/rs.3.rs-10884117/v1 — Metraipolis archive, https://doi.org/10.21203/rs.3.rs-10884117/v1 (CC-BY-4.0); archived copy: https://metraipolis.com/documento/dc_94f86c8b6441 (Metraipolis archive dc_94f86c8b6441, sha256 4d67b3ddb86ce9eef6dc8510bb8b8ed2216e00133268bb84e66355012091a161)
BibTeX
@misc{dc94f86c8b6441,
  title = {Deterministic Fact Validation for Retrieval-Augmented Generation of Clinical Data Reviewer's Guides: A Proposed Architecture (2026) · DOI 10.21203/rs.3.rs-10884117/v1},
  howpublished = {Metraipolis archive},
  url = {https://doi.org/10.21203/rs.3.rs-10884117/v1},
  note = {License CC-BY-4.0; archived copy https://metraipolis.com/documento/dc_94f86c8b6441; sha256 4d67b3ddb86ce9eef6dc8510bb8b8ed2216e00133268bb84e66355012091a161},
  year = {2026}
}

Archived text

Reproduced under CC-BY-4.0. Attribution: Metraipolis archive, https://doi.org/10.21203/rs.3.rs-10884117/v1. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Articolo recente con licenza aperta: scheda dai metadati Crossref (posted-content).
Titolo: Deterministic Fact Validation for Retrieval-Augmented Generation of Clinical Data Reviewer's Guides: A Proposed Architecture
Autori: Sai Krishna Puli
Pubblicato: 2026-09-29 · editore: Springer Science and Business Media LLC
DOI: https://doi.org/10.21203/rs.3.rs-10884117/v1
Licenza dell'articolo: CC-BY-4.0

Abstract (testo degli autori, CC-BY-4.0):
Abstract Objective: Data Reviewer's Guides (DRGs) accompanying CDISC-compliant clinical trial datasets are increasingly drafted using retrieval-augmented generation (RAG). The most detailed published account of this practice reports human review as the only safeguard between a generated draft and a submitted document, with no automated step checking generated claims against the study's own structured metadata before a reviewer opens the file. This paper proposes an architecture to close that gap. Methods: We propose a deterministic, non-large-language-model validation layer inserted between draft generation and human review, organized around a fact contract that converts each checkable clinical assertion, such as a subject count or treatment-arm allocation, into a claim-source-method-result tuple resolved by direct lookup against source metadata rather than by a second model's judgment. Because no such system has yet been implemented or deployed, we conducted a disclosed, parameterized simulation of 2,000 documents per condition comparing four pipeline configurations: a baseline LLM, retrieval-augmented generation (RAG), RAG with LLM-based self-verification, and RAG with the proposed deterministic validator, across eleven metrics spanning factual accuracy, validation coverage, reviewer workload, and latency. Results: Self-verification achieved near-complete validation coverage (1.000) but recovered only 48.0 percent of true errors among the claims it checked, with an 8.1 percent false-positive rate. The deterministic validator covered a narrower 58.7 percent of claims but recovered 99.5 percent of errors within that coverage, with a 0.5 percent false-positive rate, and produced both the highest reviewer correction rate (59.2 percent versus 44.6 percent for self-verification) and the highest final document accuracy (34.6 percent versus 16.1 percent), despite requiring more reviewer time per document (10.0 versus 5.3 minutes). Conclusion: These results are provisional by construction, since no configuration reflects a live pipeline or real reviewers, and are limited by parser brittleness, dependence on source-metadata currency, and the risk that a nominally verified claim quietly lowers reviewer vigilance elsewhere. The findings nonetheless indicate that full validation coverage can mask an automation-complacency cost that a narrower, deterministic layer avoids. A concrete empirical evaluation protocol using real documents and reviewers is proposed as the necessary next step.

Fonte: https://api.crossref.org/works/10.21203%2Frs.3.rs-10884117%2Fv1 — metadati Crossref, consultati il 2026-10-06. Interrogazione: «retrieval augmented generation».